Match each RAGAS metric to the specific RAG quality problem it detects
Drag each answer to line up with its matching prompt
Faithfulness
Retriever fetches noisy, irrelevant chunks that fill context with distractors
Answer relevance
Generator produces a grounded but off topic response that doesn't address the question
Context precision
Retriever misses important source chunks, leaving the generator without key evidence
Context recall
Generator adds claims not supported by retrieved context (hallucination)
RAGAS splits into two pairs: answer-side (faithfulness catches hallucination, answer relevance catches off-topic answers) and context-side (precision catches noisy retrieval, recall catches missing evidence).
Imagine an open-book exam where a student answers using a textbook. Four things can go wrong, and they split into two halves. On the writing side: the student might write things the book never said (a faithfulness problem), or write a perfectly book-backed paragraph that ignores the actual question (an answer-relevance problem). On the book-fetching side: the helper who hands over pages might grab lots of irrelevant pages (a precision problem), or forget to hand over the one page that actually contains the answer (a recall problem). RAGAS gives each of these four failures its own score, so when the exam goes badly you know exactly which person to blame, the writer or the page-fetcher.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
RAGAS is the landmark framework for evaluating retrieval-augmented generation without hand-written reference answers for every query. Its core design insight is that a RAG pipeline has two distinct components that fail in distinct ways, and a single quality score hides which one broke. So RAGAS splits evaluation into a clean 2x2: two metrics for the generator (the answer side) and two for the retriever (the context side).
That split is the whole point of the framework. Before RAGAS, teams mostly graded RAG end to end, with a human or an LLM judge reading the final answer and rating it good or bad. That tells you the system is broken but not which half broke, so you cannot prioritise work. By giving the retriever and the generator separate scores, RAGAS converts a vague quality complaint into a precise, actionable diagnosis.
This question asks you to match each of the four metrics to the specific failure mode it detects. The trick is to first sort the metrics into their two groups, then match within each group. Once you have the framing, the mapping is almost mechanical. The deep dive below walks each metric, how it is actually computed, why the four are pairwise orthogonal, and how to read the grid to debug a real RAG regression.
The 2x2: answer-side versus context-side
Every RAG query flows through two stages. A retriever fetches chunks from a knowledge base, and a generator writes an answer conditioned on those chunks. A failure can originate in either stage, and confusing the two is the most expensive mistake in RAG debugging.
RAGAS gives each stage its own pair of metrics. The answer-side pair, faithfulness and answer relevance, judges what the generator produced. The context-side pair, context precision and context recall, judges what the retriever fetched. This is the single framing that unlocks the whole question.
The practical payoff is attribution. If your end-to-end quality drops, you do not guess. You look at which metric moved. A drop on the context side means the retriever is the culprit and no amount of prompt engineering will save you. A drop on the answer side with healthy context means the generator is mishandling perfectly good evidence. The 2x2 turns a vague 'the answers got worse' into a precise 'the retriever stopped fetching the right chunks'.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- RAGAS scores faithfulness and answer relevance with a pinned judge like GPT-5.5 or Claude Opus 4.7, decomposing answers into atomic claims for entailment.
- LangSmith and TruLens both expose the same retriever versus generator split (TruLens calls faithfulness 'groundedness') in their RAG evaluation dashboards.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does context recall causally cap both faithfulness and answer relevance?
If the evidence was never retrieved, the generator has nothing true to ground on. Trace the dependency: recall gates what context exists, and faithfulness and relevance are both computed against that context. Fixing the generator cannot recover evidence the retriever dropped.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing context precision with context recall, or conflating faithfulness with answer relevance. They measure orthogonal failures, so swapping them sends you debugging the wrong half of the pipeline.
60 second bullets to scan on the way to the call.
The answer-side versus context-side 2x2 split
Which two metrics judge the generator
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.